DiagnoseNET is a modular framework for managing deep learning workflows, designed to simplify building and fine-tuning neural architectures. Its runtime abstracts the distributed orchestration for portability and scalability, from a GPU workstation to multi-node computational platforms. It automates neural architecture definition, hyperparameter search, data locality, and batching through a single API. The runtime coordinates parameter distribution across devices using synchronous or asynchronous gradient computations with gRPC or MPI protocols, compatible with x86 and ARM architectures.

DiagnoseNET aims to create efficient medical diagnostic tools with minimal infrastructure and low power consumption. The first application automated unsupervised patient phenotype representation using a mini-cluster of Nvidia Jetson TX2.

This workflow consists of three stages:

1. Extracting patient features from electronic health records and serializing each record into a binary format.
2. Embedding the binary matrix via unsupervised learning to identify latent phenotypic representations.
3. Using these features for supervised learning with machine learning algorithms or as initializers for deep neural networks.